What has actually been done

The site's form contract is name, plus phone or email. The n8n validation node required all three. A lead who filled in a name and an email and no phone number was rejected at the confirm button, and nobody saw it, because the rejection happened server side after the visitor had already done the work.

This had been live since 13 August. It is the reason the whole service exists. The validation rule now lives in exactly one place, frozen as data, with a test that reads the site's own form component and fails the build if the two ever drift apart again.

The shared environment file on the VM is used by every AIOS service on that machine, and it already carried a DATABASE_URL for a different project pointing at a remote Neon database. That generic name sat first in the lookup order, so the funnel service connected to the wrong database as the wrong user.

The dangerous part was the health check. It reported db: true, because a database answered. Only an actual query failed. Flipping live traffic onto that green tick would have taken the booking page dark. Caught before cutover. Generic environment names are now banned from database resolution outright.

The test suite was mutation tested: five deliberate regressions injected to see whether the tests would notice. Four were caught. The fifth, hardcoding UTC+3 instead of the Asia/Jerusalem time zone, survived all 314 tests, because in August Israel daylight time is UTC+3.

Israel returns to winter time on 25 October, and the discovery call offers slots ten working days ahead. So from the middle of October it would have started writing wrong instants into the database, wrong times into every confirmation email and wrong times into every calendar invite, silently, with nothing failing. Fourteen daylight saving tests were added and the same mutation now fails six of them.

Both emails sent correctly. The WhatsApp notification died on a name resolution error. The Evolution API address in the shared environment is a Docker internal host name, which is correct for n8n because n8n is a container, and unreachable from the new service because it runs as a host process. Fixed with a service specific override, leaving the shared value alone so the containers keep working.

Tom caught it on the live site and asked whether a full English quality pass had ever been run. It had not. Every previous round checked structure, English and Hebrew parity, accessibility, link health, console errors and word counts. Nobody had ever simply read the English as an English reader, and the agents that wrote the copy were the same ones grading it.

It got worse before it got better. The first sweep only touched one file and reported "seven down to zero" while four instances survived elsewhere, including one on step 1 of the booking page in the default path. That is a false all clear, which is worse than no check. The rule in the file header that instructed agents to use the phrase was also corrected, which is the actual reason it kept coming back.

Other things that pass a structural check and fail a reading: "points at" instead of "points to" three times, one of them a heading. A 21 word gap between subject and verb in the viability lede. "On your machines", which quietly contradicts the one person scope that the price depends on. Retired call names still sitting in every confirmation email.

The booking page had been rebuilt on 13 August around two animated characters traced out of the emblem. A critic reviewed a well known reference loader, measured that it separates its parts across ten distinct hues, and prescribed the same treatment here. That was an accurate reading of the reference and the wrong prescription for this brand, and it was applied literally instead of being questioned.

"The octopus is purple and weird, the other one looks like a Cyanide and Happiness character, both look ridiculous."

He was right. The emblem is a stark black on yellow noir mark, and a purple mantle, teal sleeves, an orange belt and auburn hair destroyed the one thing that made it good. The colour experiment was reverted. Separation now comes from value and from white, not from hue, which is how the reference actually reads too.

After the sixth rejection Tom said to remove the animation and let the logo do the work instead. The two character scene is gone from the booking page: the stage unmounted, the state machine deleted, about 170 lines of CSS removed. Animating the mark as a whole keeps what made it good. The rigged source files are preserved in case an animator is ever hired.

Two useful things survived the revert, because they were right regardless: the characters had been display:none below 1024px, meaning nobody on a phone had ever seen them, and the opacity range was too flat to read.

An invented citation, an unverifiable one, two broken internal cross references, and false factual claims about data processing agreements that were never actually executed. All removed. This is exactly what the two adversarial review passes were for.

The original fix specification named the post booking landing screen as the checkout target. It is not a checkout. octomonic.com has no payment page at all. The legal text is finished and correct and cannot go live until there is something for it to attach to, which loops back to Grow verification.

The success screen read "Confirmed. The time is taken and the calendar is updated." Behind it, the backend only wrote a 24 hour hold and emailed Tom approve and decline links. No calendar event was ever created. If Tom declined, the lead was told nothing at all.

Fixed structurally: the calendar event is now created inside the request, and the response carries an explicit confirmed or pending status, so the front end physically cannot claim a confirmation it did not receive. If the calendar call fails it degrades to a real hold with honest copy.

Slot claiming is now atomic, with a database level exclusion constraint over the time range. The losing request gets a clean "slot taken" response and an availability refresh instead of a 500.

Decline is now routed into cancel, which actually emails the lead. Both old webhook paths were kept alive so approve and decline links already sitting in Tom's inbox still answer.

The booking page mobile Lighthouse score was never 100 and was not claimed to be. It is the one route that cannot be prerendered because it needs live slot data. The animation cost 2.3KB gzipped on every page and 18KB on the booking page.

Every one of those links, pasted into LinkedIn, WhatsApp or Slack, unfurled as a bare grey card with no image, no title and no description. These are precisely the assets most meant to be shared: the payload of the email gate, the links dropped into founder groups, the featured items on a LinkedIn profile. Found by accident while writing the LinkedIn plan.

Tags added. Each demo now uses its own preview image; the sample uses the brand card.

Clean URLs turn the .html form into a 301, so every visitor to that link took an extra hop for no reason. Repointed at the canonical address.

The recommended LinkedIn headline led with the restaurant background, which Tom had explicitly killed for the website on 5 August. That was a briefing failure on the AIOS side. Both variants were shipped side by side in the document rather than quietly picking one.

Flagged as urgent by the audit, because a stale date on a proof block undermines the exact thing the proof block exists to do. The month is now stamped at build time and rendered per locale, so the copy can never lag a deploy again. It reads August 2026 because the site was built in August.

Portfolio creation does have an API. It had been reported as permanently interface only. It is actually blocked by the app's access tier, which is a different problem with a different solution. Also learned: ad account creation needs six parameters and there is a lifetime cap of five per user, and standard versus advanced access are two independent gates rather than one ladder.

The person schema omits fields when they are unset, the avatar slot renders nothing, and the acknowledgment email template has no consumer. Nothing renders or sends while empty. The round overran its one hour box and took about two.

Six questions (what is this, what do I get, what happens next, what does it cost, why trust it, what do I do now) must all be answerable from 375 pixel wide screenshots alone. If they cannot be, the page has failed regardless of how it reads on a desktop.

The WhatsApp floating button clips the closing call to action arrow at certain mobile scroll positions. It is still tappable and it is still cosmetic, and it is still on the list three weeks later.

The root element in the delivered HTML was empty. Everything a visitor saw was painted by JavaScript after the fact, which meant 82% of the time to largest contentful paint was pure render delay. On a mobile connection that is 7.8 seconds of nothing.

Real rendered HTML now goes into the static shells at build time and the application CSS is inlined, so there is no render blocking stylesheet request. Mobile Lighthouse went from 60 to 100 and largest contentful paint from 7.8 seconds to around 1.3.

A price that omits VAT in Israel is not a rounding issue, it is a number the customer will not recognise on the invoice. Critical, fixed everywhere the figure renders.

"5 to 10 minutes" came out as "10 to 5" in Hebrew, on every button on the site, because the bidirectional text algorithm was left to sort out mixed numerals on its own. Fixed inside the shared link components so it cannot come back one button at a time, and verified by checking where the glyphs actually landed on screen rather than by reading the markup.

On every page, on every load. It has been guarded and made lazy, so it stays completely inert until a real measurement ID exists.

The fix for the Hebrew range problem had split a protected text run across separate flex items, so on a narrow phone the button broke apart mid phrase. Caught by QA before deploy and fixed in the primitives.

A 727KB JavaScript chunk split into a 201KB core plus per route pieces. Favicons cut from 314KB to 46KB. Contrast raised to AA site wide, taking accessibility from 95 to 100. Navigation renamed from "AI setup" to "Workspace setup" to match what the pricing page had always called it. Pixel disclosure added to the privacy policy before any pixel existed.

Six new rules written into the header of the file the agents read before writing any copy. This is the same mechanism that later failed on "by hand", where the rule itself was wrong and every agent dutifully obeyed it. The rules file is load bearing, which cuts both ways.

Every route at two widths in both languages, checked for zero rendered hits on Israel, the revenue gate, headcount, first person and claimed tier, outside of legal and timezone context where they are legitimate.

"It shouldn't say 'I build you' or I anything. I don't want it in first person. Make sure it's corrected everywhere."

First person went from 94 hits to 7 in English and 57 to 4 in Hebrew. The survivors are the WhatsApp prefill strings, which are the visitor's own outgoing message, so rewriting them would have been wrong. Nothing first person renders in visible page text on any route.

The cause was a copy rule that said, in writing, "First person singular. OctoMonic is one person; never we/our team." Every agent had followed it correctly. The rule was inverted in place with the reversal dated and the reasoning written out, so the next agent does not restore the old wording. Recasting was preferred over a reflexive swap to "we", so the page does not read as agency boilerplate.

All eleven cards were rendering their complete desktop content stacked vertically. The detail now collapses behind a tap below tablet width, hidden with CSS rather than removed, so it stays in the document for crawlers, browser search and screen readers. Desktop is byte for byte unchanged. Down to 8,243 pixels.

The slot picker was rendering 37 pixel targets in a dense wrap. Zero interactive elements are now under 44 pixels on phone or tablet. The breakpoint for this deliberately sits where the navigation bar itself switches to desktop, so a 768 pixel tablet is treated as the finger device it actually is.

The animated system graph behind the hero was desktop only. The most designed moment on the site was invisible on the device most prospects arrive on. It now renders under the headline on phone and tablet.

The contact and workspace enquiry grids were splitting into two columns at tablet width, giving the form a 350 pixel column with 165 pixel input fields sitting beside several hundred pixels of dead space. They now split one breakpoint later.

A quarter of a 375 pixel viewport was being spent on vertical padding designed for a 1280 pixel screen.

It was displaying 2,484 and climbing. It was not measuring anything. It was a function that added a random number on a timer. Deleted, along with a hardcoded "4+ Key Industries" and a "VERSION 2.1.0 / ALL SYSTEMS NOMINAL" status bar that was also connected to nothing.

Removed. Only health checked links ship now. A retired data file carrying client identifying strings was also deleted, along with a dangling clause about it still sitting in the terms page.

Language was a runtime toggle sitting on one set of URLs. Whichever language Google's own request happened to resolve to was the only version it could index. Both languages now have real, separate, indexable URLs.

An old favicon file was winning priority over the emblem, which is why the tab looked generic and off brand. Regenerated from the actual mark.

The stylesheet pointed at Inter and Inter was never fetched, so every visitor was reading a system fallback. Both the Latin and the Hebrew faces now genuinely load.

Removed globally. Hebrew now drops uppercase, tracking and synthetic italics everywhere, because none of the three do anything good to Hebrew letterforms.

Numeric runs were being left to the bidirectional algorithm. Every figure, range and unit is now wrapped in its own isolate at render time, and it was verified by screenshot rather than by reading the markup.

A sparkline asserts a trend, and there was no trend data behind those figures. The stats got a tick meter instead: for a figure of 24 or less it draws exactly that many marks, which is a faithful encoding of the real number, and above that it draws a texture that claims nothing. The count up component treats null as its resting state and renders the source string verbatim, so every failure path lands on the true figure rather than a wrong or stuck one.